Supplier Audit System: Citation and Traceability

Data Characteristics

Supplier audit system data in the biopharmaceutical industry comes primarily from internal Quality Management System (QMS) documents. Examples include "Supplier Quality Management Procedures," "Approved Supplier List," and "Audit Checklists." These documents are typically stored as PDFs, Word files, or Excel spreadsheets. Update frequency depends on supplier risk levels and internal management requirements, usually annually or biennially. High-risk suppliers may require more frequent updates.

Document structures often include extensive normative text, flowcharts, approval records, and key parameter tables. Fields typically cover supplier qualifications, production capacity, quality system certifications, historical audit results, non-conformances, and Corrective and Preventive Actions (CAPA). Units often involve production volume (e.g., kg/batch), purity (e.g., %), and deviation values (e.g., ppm). Traceability information such as batch numbers and expiration dates is also common.

Constraints on Citation and Traceability

The normative and traceability requirements of supplier audit data impose specific constraints on the citation and traceability process. Flowcharts and table data within documents require the knowledge base to accurately parse and link them to specific paragraphs, preventing information loss.

Frequent updates mean the knowledge base must support version management and incremental updates. This ensures that citations always refer to the latest version of the system. Specific fields and units, such as Lot No. or mg/ml, commonly found in audit reports, require accurate processing by the tokenizer and entity recognition capabilities to support precise retrieval.

Audit content involves compliance, so users demand high accuracy and traceability for citation sources. Fabricated or inaccurate citations can lead to severe compliance risks. The system must ensure each answer links to the exact location in the original document and avoid generating citation markers without a source.

Configuration Guidelines

Configuration Item	Suggested Value	Rationale
`Chunk size` (Segment Length)	`500–800 characters`	Supplier audit document paragraphs are typically long and logically complete. Shorter segments would break context.
`Recall count` (Retrieval Count)	`top 8–12 items`	Audit systems are complex. More contextual information is needed to support accurate answers.
`Similarity threshold` (Similarity Threshold)	`0.8–0.85`	Audit content is rigorous. A higher similarity is needed to ensure retrieval accuracy and reduce miscitations.
`Rerank result count` (Reranked Return Count)	`top 5 items`	Reduces noise for the model, focusing on the most relevant citations.
`Citation Display Mode`	`Show Citations Only`	Forces the model to output only sourced citations, preventing false citations.
`ENABLE_CITATION`	`true`	Ensures the system enforces citation marking when generating responses.

Common Pitfalls

Model responses show citation Marker：[1] (Citation Marker: [1]) or citation IDs do not match the original text: This occurs when the model fails to correctly parse or filter internal citation markers during generation. Review the model's post-processing logic.
Answer content is inconsistent with the citation source, or references non-existent document IDs: This usually indicates an outdated knowledge base index or model "hallucinations" regarding citations. Strengthen the knowledge base's real-time synchronization mechanism and the model's understanding of the citation mechanism.
Uploaded audit reports (e.g., PDF format) fail to parse or have missing content: This may be due to complex document structures, numerous images, or scanned documents. The PARSE_FILE_TIMEOUT_SECONDS setting might be too short, or the parser version might be too old.

Verification Steps

Select a typical supplier audit process document. Ask questions about key steps or responsible parties. Verify that answers are accurate and that each citation marker links to the corresponding location in the original document.
Upload an updated supplier qualification file. Wait for the knowledge base to index it. Then, query the latest qualification information for that supplier. Confirm that the citation source has been updated to the latest version.
Ask questions about specific batch numbers or product specifications from an audit report. Verify that the answer accurately identifies and cites paragraphs containing these key fields, and that units (e.g., mg/ml) are correct.

Note: The values provided are common starting points. Measure them against your own samples for optimal performance.

Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.